John T. Sheehan wrote:
Thanks for the tutorial. I will ask Product Management if the stripping of the units is intentional or an oversight in regression. Thanks
I could image that the problem is - stripping which units. The numerical algorithms use absolute values to decide if we have a convergence and if precision is good enough. That way it really matters if for volume Liters are stripped or m^3. So, should the algorithm strip the base (SI) units or the units the user is specifying when defining the data and guess values (are they still available?)? So question is if its possible to find a meaningful and intuitive way to deal with units here - am not sure about that. On the other hand it would be nice if a model would be refused right at the beginning because of obvious unit mismatch. E.g. let t be time and V volume as in the example in this thread. A model like g(t,v):=v[0 * ln(v[0 * t + v[1) would be refused because v[0 can't be volume and time^-1 at the same time.
And a built in standard regression like logfit() with the form a * ln(x + b) + c would not be possible at all with data with units, as x has to be assumed unitless here unless we provide an additonal coefficient for x (which we could normalize to 1/UnitOf(x)??). So it seems that allowing units in regressions would require some rewriting of predefined egressions as well.
